Research map / Large language models / Retrieval and knowledge
PatternBloom: Empowering Agentic RAG with Externalized RL-Distilled Graph Patterns
Yi Ge, Heng Huang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PatternBloom, which externalizes reasoning routes over retrieved graph content as reusable patterns distilled via RL for agentic RAG.
Problem
RL-trained agentic RAG systems conflate declarative knowledge (retrieved content) with procedural knowledge (reasoning routes recurring across queries), since the reasoning route stays inside the LLM's hidden state.
Approach
Externalizes the reasoning route through retrieved graph content as graph patterns distilled from RL-trained agents, separating procedural from declarative knowledge; the abstract only partially describes the method.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add PAR pattern-coverage reward (Stage II) (ablation) | PatternBloom 7B · HotpotQA/2Wiki/MuSiQue/Bamboogle avg | Avg-F1: 44.1 -> 55.9 | Improved |
| add Graph Pattern Memory (ablation) | PatternBloom 7B · 4 multi-hop avg | Avg-F1: 46.6 -> 55.9 | Improved |
| add IDR Stage I pre-training (ablation) | PatternBloom 7B · 4 multi-hop avg | Avg-F1: 52.9 -> 55.9 | Improved |
| set shaping weight lambda_p to 0.2 rather than 0.1 (ablation) | PAR · 4 multi-hop avg | Avg-F1: 45.2 -> 55.9 | Improved |
| set shaping weight lambda_p to 0.2 rather than 0.5 (ablation) | PAR · 4 multi-hop avg | Avg-F1: 43.3 -> 55.9 | Improved |
| use full 27-pattern pool instead of top-1 (ablation) | GPM retrieval · 4 multi-hop avg | Avg-F1: 43.0 -> 55.9 | Improved |
| add Stage I (IDR) vs Stage II removal on Bamboogle (ablation) | PatternBloom 7B · Bamboogle | F1: 61.2 vs 66.0 (no PAR), 68.8 (no GPM) | Mixed |
| use PatternBloom instead of Search-R1 7B | agentic RAG 7B · Avg-OOD | EM: 38.1 -> 49.0 | Improved |
| use PatternBloom instead of Structure-R1 on 2Wiki | agentic RAG 7B · 2WikiMultiHopQA | EM: 69.2 -> 54.0 | Hurt |
| add pattern memory to reduce turns vs Graph-R1 | agentic RAG · Bridge questions | avg turns: 5.4 -> 4.7 | Improved |
| scale PatternBloom backbone 7B to 32B | PatternBloom · Avg-OOD | EM: 49.0 -> 52.8 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Graph-R1 — Graph-R1: Towards Agentic GraphRAG Framework via End-to-end Reinforcement Learning (ICML 2026) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents (ICML 2026 · similarity 0.89)
- MAGE: Multi-Agent Self-Evolution with Co-Evolutionary Knowledge Graphs (NeurIPS 2026 · similarity 0.88)
- MC-Search: Evaluating and Enhancing Multimodal Agentic Search with Structured Long Reasoning Chains (ICLR 2026 · similarity 0.88)
- SD-Search: Hindsight Self-Distillation for Search-Augmented Reasoning (NeurIPS 2026 · similarity 0.88)
- ReSeek: A Self-Correcting Framework for Search Agents with Instructive Rewards (ICML 2026 · similarity 0.88)
- Search-R2: Enhancing Search-Integrated Reasoning via Actor-Refiner Collaboration (ICML 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
